Traffic flow prediction model based on improved fully connected neural network

ObjectiveIn response to the problem of unsatisfactory prediction accuracy of existing models, an improved fully connected neural network model was proposed to further improve traffic flow prediction accuracy.MethodFirst, defined a custom layer, where added a weight matrix and a bias term. Second, input a linear transformation, which multiplied the weight and added a bias term; and then performed nonlinear transformation through activation function for forward propagation. Third, customized a model function named Explainer_model, which was used to build, train and predict models. A flattening layer was adopted to expand input data into a one-dimensional array. Fourth, the Dropout layer between two fully connected layers was used to reduce the risk of overfitting. The fully connected neural networks were trained by using backpropagation algorithms. Nadam optimizer with the best optimization effect was used, in terms of optimizer selection after comparative verification. Finally, addd a fully connected layer with one output unit for outputting the results.ResultThe experiments were conducted using hourly traffic flow data at four collection points on a certain expressway in Shaanxi, China to verify the effectiveness of model. The improved neural network was used to model the time series and compared with four neural network models, i.e., GRU, LSTM, CNN-LSTM and CNN-GRU. The result indicates that the improved fully connected neural network model performs better than other models.ConclusionThe proposed predictive model can be used to predict data with seasonality and trends.